Low-light Image Enhancement Using Deep Neural Network: An Improvement on ZeroDCE++

This paper presents a no-reference deep neural network technique to improve the quality of images captured in challenging lighting conditions that can overcome issues like noise and loss of object detail. Several effective methods, like ZeroDCE++, enhance low-light images. However, most perform well only on images similar to those seen during training, i.e. images under a narrow range of brightness variation. In this work, first a pre-processing step is designed to mitigate noise in dark areas of low-light images and to prevent saturation in regions with excessive illumination. In addition, a loss function is introduced to preserve color balance during the training phase. The qualitative and quantitative results show that the proposed method significantly enhances low-light images in comparison with SOTA especially with ZeroDCE++.

کلیدواژه ها:

Low-light image ، Image enhancement ، Deep neural network ، No-reference method